forgot to threshold exemplar

app.py

@@ -72,6 +72,7 @@ def load_fn(images, rng_state, exemplars_state, lut_state):
     exemplars_state[0] = rs.binomial(n=1, p=0.5, size=HYPERDIMS)
     exemplars_state[1] = rs.binomial(n=1, p=0.5, size=HYPERDIMS)
     lut_state[0] = make_LUT(2**VALUE_BITS, HYPERDIMS, rs)
+    assert lut_state[0].shape[0] == val_bins.shape[0]
     lut_state[1] = make_LUT(2**POS_BITS, HYPERDIMS, rs)
     print(exemplars_state)
     print(lut_state[0].shape, lut_state[1].shape)
@@ -87,7 +88,7 @@ def quantize_embeds(embeds):
     quantized_embeds_flat = val_bins[closest_bin]
     quantized_embeds = np.reshape(quantized_embeds_flat, embeds.shape)
     closest_bin = np.reshape(closest_bin, embeds.shape)
-    print(closest_bin.shape)
+    print(closest_bin.shape, 'values are in bins', closest_bin.min(), 'to', closest_bin.max())
     return quantized_embeds, closest_bin
 
 def update_exemplars(df, rng, exemplars, lut):
@@ -129,6 +130,10 @@ def update_exemplars(df, rng, exemplars, lut):
     exemplars_integer[1] = np.sum(hd_embeds[labels_train == 1], axis=0)
     exemplars[0] = exemplars_integer[0] / np.sum(labels_train == 0)
     exemplars[1] = exemplars_integer[1] / np.sum(labels_train == 1)
+    exemplars[0][exemplars[0] >= 0.5] = 1.
+    exemplars[0][exemplars[0] < 0.5] = 0.
+    exemplars[1][exemplars[1] >= 0.5] = 1.
+    exemplars[1][exemplars[1] < 0.5] = 0.
     print(exemplars[0].shape, exemplars[1].shape, np.abs(exemplars[0] - exemplars[1]).sum())
     
     return rng, exemplars, 0., 0. # score(embeds_train, exemplars, lut), score(embeds_test, exemplars, lut)